app.py

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import gradio as gr

# Load the base model
base_model_name = "unsloth/llama-3.2-3b-instruct-bnb-4bit"
tokenizer = AutoTokenizer.from_pretrained(base_model_name, use_fast=False)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    device_map="auto",  # Automatically map layers to available devices
    torch_dtype=torch.float16  # Ensure compatibility with 4-bit quantization
)

# Load the LoRA adapter
adapter_path = "Grandediw/lora_model"  # Replace with your model path
model = PeftModel.from_pretrained(base_model, adapter_path)
model.eval()  # Set the model to evaluation mode

# Define the inference function
def respond(
    message,
    history: list[tuple[str, str]],
    max_tokens,
    temperature,
    top_p,
):
    # Build context from history
    context = ""
    for user_message, assistant_message in history:
        context += f"User: {user_message}\nAssistant: {assistant_message}\n"
    context += f"User: {message}\nAssistant:"

    # Tokenize the input
    inputs = tokenizer(context, return_tensors="pt").to("cuda")

    # Generate a response
    outputs = model.generate(
        input_ids=inputs.input_ids,
        max_new_tokens=max_tokens,
        temperature=temperature,
        top_p=top_p,
        do_sample=True
    )

    # Decode and return the response
    response = tokenizer.decode(outputs[:, inputs.input_ids.shape[-1]:][0], skip_special_tokens=True)
    return response

# Build the Gradio ChatInterface
demo = gr.ChatInterface(
    fn=respond,
    additional_inputs=[
        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max Tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=1.5, step=0.1, label="Temperature"),
        gr.Slider(minimum=0.1, maximum=1.0, value=0.9, step=0.05, label="Top-p"),
    ],
)

if __name__ == "__main__":
    demo.launch()